subtraction

The passage around this formula

Put simply: if a search evaluates k candidate directions against a validation check, and each candidate independently has a small probability $\alpha$ of appearing to pass that check purely by chance — because it happens to correlate with the labeled examples used to test it — then the probability that the search turns up at least one apparent “finding” is $P_{\text{at least one}} = 1 - (1-\alpha)^k$. a quantity that climbs toward near-certainty long before k reaches even a small fraction of the directions a real residual stream offers. A single successful-looking direction, reported without the number of candidates it was selected from, is close to uninterpretable on its own.

Read this part in the article →